Statistical methods from machine learning have been key to the progress of computer vision in recent years. Use of machine learning has led to development of many successful vision applications such as face and pedestrian detectors. Along the same principles of using robust statistical methods, in this thesis we build systems for two biomedical imaging domains.

The first system detects cancer in prostate pathology images. Recent technological advancements have made it possible to commercially build whole-slide scanning microscopes that generate digital images of whole-slides at magnifications required for an effective clinical diagnosis. The availability of digital images of pathology slides allows development of Computer Aided Diagnostic (CAD) tools that can improve pathologist's accuracy and efficiency in diagnosis. An automated screener can assist pathologists in diagnosing by suggesting suspicious locations. The screening tool can also reduce the bandwidth required for diagnosing remotely by transferring only the suspicious parts. To provide a base on which such CAD tools can be developed, we build a cancer detector for prostate needle core biopsies, which is one of the most frequently diagnosed tissue.

The second system analyzes High-Throughput Screening (HTS) images of C. elegans worms to identify their phenotype. HTS is a class of biological experiments where a large number of similar experiments are conducted to identify a small number of drugs or genes relevant to a biological process. Recently, researchers have started conducting HTS experiments using C. elegans in which the experimental output are images. To assist biologists in analyzing the large number of images generated by the HTS experiments on C. elegans, we develop a system that identifies a worm's phenotype. A preferred way of conducting such experiments is to image the worms in agar. The shadows cast by track marks left by the worms in agar appear similar to the worms which complicates segmenting images of worms. To reliably segment worms in such conditions, we develop a novel segmentation technique that uses multiple visual cues such as texture, contrast and shape to segment the worms. After segmenting the worms, our system also analyzes the fluorescent patterns inside the worms to identify their phenotype.